New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

squeezenet bug #6500

Closed
fradino opened this Issue Sep 30, 2018 · 5 comments

Comments

Projects
None yet
3 participants
@fradino
Copy link

fradino commented Sep 30, 2018

Exception in thread "main" java.lang.IllegalStateException: Invalid configuration: Vertex "fire2_relu_exp1x1" has input "fire2_exp1x1" that does not exist
at org.deeplearning4j.nn.conf.ComputationGraphConfiguration.validate(ComputationGraphConfiguration.java:314)
at org.deeplearning4j.nn.conf.ComputationGraphConfiguration$GraphBuilder.build(ComputationGraphConfiguration.java:1033)
at org.deeplearning4j.zoo.model.SqueezeNet.init(SqueezeNet.java:78)
at org.deeplearning4j.zoo.model.SqueezeNet.init(SqueezeNet.java:36)

@Charele

This comment has been minimized.

Copy link

Charele commented Sep 30, 2018

I check the code, some typos here:
1>.addLayer(prefix+"exp1x1", new ConvolutionLayer.Builder(1, 1).nOut(expand)
.cudnnAlgoMode(cudnnAlgoMode).build(), prefix+"_relu_sq1x1")
the "exp1x1" shoud be "_exp1x1"

2> .addLayer("conv10", new ConvolutionLayer.Builder(1,1).nOut(numClasses)
.cudnnAlgoMode(cudnnAlgoMode).build(), "input")
I think the "conv10" should connect to "drop9"?

3>
.addLayer(prefix+"exp1x1", new ConvolutionLayer.Builder(1, 1).nOut(expand)
.cudnnAlgoMode(cudnnAlgoMode).build(), prefix+"_relu_sq1x1")
.addLayer(prefix+"_relu_exp1x1", new ActivationLayer(Activation.RELU), prefix+"_exp1x1")

            .addLayer(prefix+"_exp3x3", new ConvolutionLayer.Builder(3,3).nOut(expand)
                    .cudnnAlgoMode(cudnnAlgoMode).build(), prefix+"_relu_sq1x1")
            .addLayer(prefix+"_relu_exp3x3", new ActivationLayer(Activation.RELU), prefix+"_exp3x3")

            .addVertex(prefix, new MergeVertex(), prefix+"_relu_exp1x1", prefix+"_relu_exp3x3");

The "..._relu_exp1x1" and "..._relu_exp3x3" have different shape, they cann't be merge,
I think above two new ConvolutionLayer.Builder() should have same kernelSize,,,my option

@AlexDBlack

This comment has been minimized.

Copy link
Member

AlexDBlack commented Sep 30, 2018

As per here: #6501 (comment)

Again, try on 1.0.0-beta2.
And if it still occurs, please give us some code to reproduce this. We can't do much without knowing exactly what you are doing.

@fradino

This comment has been minimized.

Copy link

fradino commented Oct 1, 2018

now,I use 1.0.0-beta2
the code I write is just

ZooModel zooModel = SqueezeNet.builder().build();
ComputationGraph sqnet = zooModel.init();
log.info(sqnet.summary());

still has this problem

AlexDBlack added a commit that referenced this issue Oct 1, 2018

@AlexDBlack AlexDBlack referenced this issue Oct 1, 2018

Merged

Misc fixes #6504

@AlexDBlack

This comment has been minimized.

Copy link
Member

AlexDBlack commented Oct 1, 2018

Fixed here: #6504

Once it's merged, you can get the fix using snapshots: https://deeplearning4j.org/docs/latest/deeplearning4j-config-snapshots

Or, in this case, you can probably just copy the updated model code to a local project (as the changes are isolated to the zoo model class) - https://deeplearning4j.org/docs/latest/deeplearning4j-config-snapshots

AlexDBlack added a commit that referenced this issue Oct 3, 2018

Misc fixes (#6504)
* #6501 Fix multiple issues with Xception configuration

* #6500 Fix SqueezeNet (non-pretrained) config + add tests

* #6497 Fix multi-output net evaluation for ComputationGraph

* #6502 fix formatting for ComputationGraph summary

* #6502 fix formatting for MultiLayerNetwork summary

* #6489 fix JSON mapping for keras import preprocessors

* Switch default output layer activation to softmax

* Fix transposei issues with autoencoder

* Minor test fixes

* ComputationGraph: Fix possible double application of preprocessors and setting input for output layers when doing backprop

* Workspace fix for output layer dropout in compgraph
@lock

This comment has been minimized.

Copy link

lock bot commented Nov 2, 2018

This thread has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs.

@lock lock bot locked and limited conversation to collaborators Nov 2, 2018

Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.